Free AI Support Report
Issue one: the best free online AI for Word Push today, our biggest problems, a stop list of 56 dead models, and the last page on services that come with strings.
Good morning! This is the very first Free AI Support Report. Every morning it tells you which free AI can help our game, in plain words, with proof for every claim.
Ten helpers and I searched today: thinking models, 3D, pictures, sound, game tools, GitHub projects and licences. A fact-checker then re-opened 70 items and fixed 88 small and big mistakes before anything was printed.
One rule runs through the whole report: only free AI that runs ONLINE. Nothing to download, nothing that heats up our PCs. Every AI in here is a service we can call from a script with a free key, so our PCs only send the question and get the answer back.
Three things stood out. One: the best free online brains are strong now, and 30 of them answered our keys this morning. Two: for pictures, 3D and voices, the free online services are fewer, and most come with daily limits or rules, so I list each catch clearly. Three: for the problems we keep fighting, like units jamming at gates and the 57 silent sounds, plain free tools (no AI at all) are often the best fix.
Today is the big build day. From tomorrow, the Factory PC collects the news and free online models do most of the writing, and I only do a final check. So the report gets cheaper every day, not more expensive.
— Claude, your editor
Every issue starts with the score: how many jobs free AI did for us since the last issue, which models did them, and which models are useless. I only print numbers that were really logged.
Counted from 2026-11-03 00:00 to 2026-11-04 23:59 UTC. 0 jobs came from the free AI logbook and 374 from the morning checks. Machines: Factory PC 358, Mini Boss 16.
| Model | Best at | Where | Jobs | Worked | Project | What the jobs were | Tokens |
|---|---|---|---|---|---|---|---|
| gemini-3-flash-preview | fast all-round reasoning and coding | gemini | 3 | 3 | free-ai-report | morning ping | — |
| deepseek-v4.1-flash-free | coding and step-by-step reasoning | opencode | 3 | 3 | free-ai-report | morning ping | — |
| openai/gpt-oss-120b | fast reasoning and tool use | groq | 4 | 4 | free-ai-report | morning ping | — |
| stepfun-ai/step-5-preview | reasoning and agent tasks | nvidia | 4 | 2 | free-ai-report | morning ping | — |
| minimaxai/minimax-m3.1-flash | fast coding and tool use | nvidia | 4 | 4 | free-ai-report | morning ping | — |
| glm-5.3-flash-free | coding and general chat | opencode | 3 | 3 | free-ai-report | morning ping | — |
| mimo-v2.6-flash-free | fast reasoning and coding | opencode | 3 | 3 | free-ai-report | morning ping | — |
| mistralai/mistral-large-4-1t-a52b | multilingual chat and writing | nvidia | 4 | 4 | free-ai-report | morning ping | — |
| gemini-3.1-flash-lite-preview | cheap fast text and pictures | gemini | 3 | 1 | free-ai-report | morning ping | — |
| gemini-3.1-flash-lite | cheap fast text and pictures | gemini | 4 | 2 | free-ai-report | morning ping | — |
| poolside/laguna-xs-2.1 | agentic coding and terminal tasks | nvidia | 4 | 4 | free-ai-report | morning ping | — |
| openai/gpt-oss-20b | very fast light reasoning | groq | 4 | 4 | free-ai-report | morning ping | — |
| qwen/qwen3.8-27b | coding and multilingual reasoning | groq | 4 | 4 | free-ai-report | morning ping | — |
| beam-free | general chat and writing | opencode | 3 | 3 | free-ai-report | morning ping | — |
| nvidia/nemotron-3.5-super-vl-120b-a12b | pictures plus text | nvidia | 3 | 3 | free-ai-report | morning ping | — |
| nemotron-3.5-super-vl-free | pictures plus text | opencode | 3 | 3 | free-ai-report | morning ping | — |
| nvidia/nemotron-3-ultra-550b-a55b | long reasoning and agents | nvidia | 4 | 4 | free-ai-report | morning ping | — |
| gemini-3.5-flash-lite | cheap fast short answers | gemini | 3 | 3 | free-ai-report | morning ping | — |
| nvidia/nemotron-3.5-lightning-30b-a3b | fast cheap short answers | nvidia | 3 | 3 | free-ai-report | morning ping | — |
| nemotron-3.5-lightning-free | fast cheap short answers | opencode | 3 | 3 | free-ai-report | morning ping | — |
| meta/muse-glimmer-30b | fast light chat | nvidia | 3 | 3 | free-ai-report | morning ping | — |
| allam-2-7b | Arabic and English chat | groq | 3 | 3 | free-ai-report | morning ping | — |
| gemini-flash-lite-latest | cheap fast short answers | gemini | 3 | 3 | free-ai-report | morning ping | — |
| gemma-4-31b-it | text, pictures, general chat | gemini | 4 | 3 | free-ai-report | morning ping | — |
| gemma-4-26b-a4b-it | fast text and pictures chat | gemini | 3 | 3 | free-ai-report | morning ping | — |
| nvidia/nemotron-3-nano-omni-30b-a3b-reasoning | text, pictures, audio understanding | nvidia | 3 | 3 | free-ai-report | morning ping | — |
| nvidia/nemotron-3.5-bolt-e4b | tiny fast simple tasks | nvidia | 3 | 3 | free-ai-report | morning ping | — |
| big-pickle | coding (stealth test model) | opencode | 3 | 3 | free-ai-report | morning ping | — |
| flint-40b-free | unknown, test model | opencode | 3 | 3 | free-ai-report | morning ping | — |
| gemini-robotics-er-2-preview | robot planning and pictures | gemini | 3 | 3 | free-ai-report | morning ping | — |
| nvidia/ising-calibration-1.5-31b | nvidia | 3 | 2 | free-ai-report | morning ping | — | |
| meta/llama-3.2-90b-vision-instruct | nvidia | 3 | 1 | free-ai-report | morning ping | — |
These models did not answer once in any check today, for a reason that will not fix itself: the service lists them but they are not there, they need a different kind of request, they are not chat models, or they are blocked in our region. From tomorrow the morning check skips them. Once a week I look at them again in case they come back.
gemini / antigravity-preview-05-2026 · gemini / antigravity-preview-09-2026 · gemini / antigravity-preview-latest
gemini / gemini-2.5-flash · gemini / gemini-2.5-flash-lite · gemini / gemini-2.5-pro · nvidia / 01-ai/yi-large · nvidia / adept/fuyu-8b · nvidia / ai21labs/jamba-1.5-large-instruct · nvidia / aisingapore/sea-lion-7b-instruct · nvidia / bigcode/starcoder2-15b · nvidia / databricks/dbrx-instruct · nvidia / deepseek-ai/deepseek-coder-6.7b-instruct · nvidia / google/codegemma-1.1-7b · nvidia / google/codegemma-7b · nvidia / google/deplot · nvidia / google/gemma-2b · nvidia / google/gemma-3-12b-it · nvidia / google/gemma-3-4b-it · nvidia / google/recurrentgemma-2b · nvidia / ibm/granite-3.0-3b-a800m-instruct · nvidia / ibm/granite-3.0-8b-instruct · nvidia / ibm/granite-34b-code-instruct · nvidia / ibm/granite-8b-code-instruct · nvidia / meta/codellama-70b · nvidia / meta/llama-3.2-11b-vision-instruct · nvidia / meta/llama2-70b · nvidia / microsoft/kosmos-2 · nvidia / microsoft/phi-3-vision-128k-instruct · nvidia / microsoft/phi-3.5-moe-instruct · nvidia / mistralai/codestral-22b-instruct-v0.1 · nvidia / mistralai/mistral-7b-instruct-v0.3 · nvidia / mistralai/mistral-large · nvidia / mistralai/mistral-large-2-instruct · nvidia / mistralai/mixtral-8x22b-v0.1 · nvidia / moonshotai/kimi-k2.6 · nvidia / nvidia/cosmos-reason2-8b · nvidia / nvidia/llama-3.1-nemotron-51b-instruct · nvidia / nvidia/llama-3.1-nemotron-70b-instruct · nvidia / nvidia/llama-3.1-nemotron-ultra-253b-v1 · nvidia / nvidia/llama3-chatqa-1.5-70b · nvidia / nvidia/mistral-nemo-minitron-8b-8k-instruct · nvidia / nvidia/nemotron-3-super-120b-a12b · nvidia / nvidia/nemotron-4-340b-instruct · nvidia / nvidia/nemotron-nano-3-30b-a3b · nvidia / nvidia/neva-22b · nvidia / nvidia/vila · nvidia / writer/palmyra-creative-122b · nvidia / writer/palmyra-fin-70b-32k · nvidia / writer/palmyra-med-70b · nvidia / writer/palmyra-med-70b-32k · nvidia / zyphra/zamba2-7b-instruct
groq / canopylabs/orpheus-arabic-saudi · groq / canopylabs/orpheus-v1-english
opencode / crouton-alpha-free · opencode / pulse-alpha-free
These failed this time for a reason that may pass, like a used-up free quota or a busy server. They get another chance tomorrow.
gemini / deep-research-max-preview-04-2026 (Free quota used up when we checked; try again tomorrow) · gemini / deep-research-preview-04-2026 (Free quota used up when we checked; try again tomorrow) · gemini / deep-research-pro-preview-12-2025 (Free quota used up when we checked; try again tomorrow) · gemini / gemini-2.5-computer-use-preview-10-2025 (Free quota used up when we checked; try again tomorrow) · gemini / gemini-3.1-flash-live-preview (Failed today: HTTP 400 [{ "error": { "code": 400, "message": "mo) · gemini / gemini-3.1-pro-preview (Free quota used up when we checked; try again tomorrow) · gemini / gemini-3.1-pro-preview-customtools (Free quota used up when we checked; try again tomorrow) · gemini / gemini-3.5-flash (Free quota used up when we checked; try again tomorrow) · gemini / gemini-3.5-live-translate-preview (Failed today: HTTP 400 [{ "error": { "code": 400, "message": "mo) · gemini / gemini-3.6-flash (Free quota used up when we checked; try again tomorrow) · gemini / gemini-3.7-flash (Service busy or broken when we checked) · gemini / gemini-3.8-flash (Free quota used up when we checked; try again tomorrow) · gemini / gemini-3.8-live (Failed today: HTTP 400 [{ "error": { "code": 400, "message": "mo) · gemini / gemini-3.8-live-extended-thinking (Failed today: HTTP 400 [{ "error": { "code": 400, "message": "mo) · gemini / gemini-4-argon (Free quota used up when we checked; try again tomorrow) · gemini / gemini-flash-latest (Service busy or broken when we checked) · gemini / gemini-nano-banana-2.1 (Free quota used up when we checked; try again tomorrow) · gemini / gemini-omni-1.1-flash (Free quota used up when we checked; try again tomorrow) · gemini / gemini-omni-flash-preview (Free quota used up when we checked; try again tomorrow) · gemini / gemini-pro-latest (Free quota used up when we checked; try again tomorrow) · gemini / gemini-robotics-er-2-streaming-preview (Failed today: HTTP 400 [{ "error": { "code": 400, "message": "mo) · gemini / lyria-3-pro-preview (Free quota used up when we checked; try again tomorrow) · gemini / lyria-3.5 (Free quota used up when we checked; try again tomorrow) · gemini / lyria-realtime-exp (Failed today: HTTP 400 [{ "error": { "code": 400, "message": "mo) · gemini / nano-banana-pro-preview (Free quota used up when we checked; try again tomorrow) · nvidia / deepseek-ai/deepseek-v4.1-flash (Too slow: no answer in time) · nvidia / google/diffusiongemma-26b-a4b-it (Failed today: empty reply) · nvidia / google/gemma-4-31b-it (Too slow: no answer in time) · nvidia / nvidia/ai-synthetic-video-detector (Service busy or broken when we checked) · nvidia / openai/gpt-oss-20b (Too slow: no answer in time) · nvidia / z-ai/glm-5.3-flash (Too slow: no answer in time) · opencode / jev-1.13-free (Failed today: Error: { "name": "UnknownError", "data": { ) · opencode / nemotron-3-ultra-free (Failed today: Error: { "name": "UnknownError", "data": { )
Three free online AI models I would pick first today. All three answered our free keys this morning, and nothing needs to be downloaded.
Gemini 3 Flash is the smartest free model that answered us this morning (top of the table below). Fast, good at code and reasoning, and it also reads pictures. The catch: on the free plan Google may read our prompts, so no secrets.
The strongest free coder we can reach. MIT license. OpenCode gives it away free online, and it answered there this morning (in 16 seconds).
Not the smartest, but the fastest good coder: it answered our free Groq key in 0.2 seconds. Great for quick jobs, tests and helpers. Apache-2.0.
Real numbers from public leaderboards. A higher score means smarter on their tests. These boards rank ALL models, paid ones too, so you can see how close the free ones are.
| Model | Board | Score or rank | Seen |
|---|---|---|---|
| Gemini 4 Carbon (high) | Artificial Analysis Intelligence Index | 60 (#1) | 2026-11-04 |
| Claude Opus 5.5 (max) | Artificial Analysis Intelligence Index | 59 (#2) | 2026-11-04 |
| Kimi K3.1 Pro (best open weights seen) | Artificial Analysis Intelligence Index | 46 | 2026-11-04 |
| Qwen3.8 Max (0902) | Artificial Analysis Intelligence Index | 34 | 2026-11-04 |
| DeepSeek V4.1 Flash 1017 (max) | Artificial Analysis Intelligence Index | 33 | 2026-11-04 |
| GLM-5.3 (max) | Artificial Analysis Intelligence Index | 32 | 2026-11-04 |
| Grok-4.1 Thinking | LMArena Text (via llm-stats mirror, only 2 rows shown) | 1483 | 2026-11-04 |
| DeepSeek V4.1 Flash | Vendor card: Terminal-Bench 2.1 | 90.6 (self-reported) | 2026-11 |
| Qwen3.8-27B | Vendor card: SWE-bench Pro | 61.7 (self-reported) | 2026-08 |
| Gemma 4 31B | Vendor card: LiveCodeBench v6 | 80.0 (self-reported) | 2026-07 |
Every morning the Factory PC and Mini Boss send each free online model a tiny message: "Reply with the single word: pong." If it answers, it is alive. It is like a roll call at school. The big table lists every model that answered, the smartest at the top.
Models are put in tiers using the Artificial Analysis Intelligence Index where it exists (v5.0.1 from the live leaderboard, v4.0 for a few older entries; the two scales differ, so the tier is set by judgement, and scores are only compared within the same version), and one coding model uses its vendor benchmark. Tier 5 means no public score was found, so those models are listed last and are not proven weak.
| # | Model | Best at | Smarts | Where | Answer time |
|---|---|---|---|---|---|
| 1 | Gemini 3 Flash Preview gemini-3-flash-preview | fast all-round reasoning and coding | Top (46 AA v4) | gemini | 0.9 s |
| 2 | DeepSeek V4.1 Flash deepseek-v4.1-flash-free | coding and step-by-step reasoning | Top (33 AA v5) | opencode | 16.3 s |
| 3 | GPT-OSS 120B openai/gpt-oss-120b | fast reasoning and tool use | Strong (33 AA v4) | groq | 0.4 s |
| 4 | Step 5 Preview stepfun-ai/step-5-preview | reasoning and agent tasks | Strong (29 AA v5) | nvidia | 22.3 s |
| 5 | MiniMax M3.1 Flash minimaxai/minimax-m3.1-flash | fast coding and tool use | Strong (29 AA v5) | nvidia | 0.6 s |
| 6 | GLM-5.3 Flash glm-5.3-flash-free | coding and general chat | Strong (28 AA v5) | opencode | 32.5 s |
| 7 | MiMo V2.6 Flash mimo-v2.6-flash-free | fast reasoning and coding | Strong (27 AA v5) | opencode | 13.8 s |
| 8 | Mistral Large 4 mistralai/mistral-large-4-1t-a52b | multilingual chat and writing | Strong (26 AA v5) | nvidia | 0.5 s |
| 9 | Gemini 3.1 Flash-Lite Preview gemini-3.1-flash-lite-preview | cheap fast text and pictures | Good (34 AA v4) | gemini | 0.7 s |
| 10 | Gemini 3.1 Flash-Lite gemini-3.1-flash-lite | cheap fast text and pictures | Good (34 AA v4) | gemini | 0.5 s |
| 11 | Laguna XS 2.1 poolside/laguna-xs-2.1 | agentic coding and terminal tasks | Good (70.9 SWE-bench %) | nvidia | 0.2 s |
| 12 | GPT-OSS 20B openai/gpt-oss-20b | very fast light reasoning | Good (24 AA v4) | groq | 0.1 s |
| 13 | Qwen3.8 27B qwen/qwen3.8-27b | coding and multilingual reasoning | Good (20 AA v5) | groq | 0.2 s |
| 14 | Beam 1.0 beam-free | general chat and writing | Good (18 AA v5) | opencode | 9.7 s |
| 15 | Nemotron 3.5 Super VL 120B nvidia/nemotron-3.5-super-vl-120b-a12b | pictures plus text | Good (14 AA v5) | nvidia | 0.8 s |
| 16 | Nemotron 3.5 Super VL (free) nemotron-3.5-super-vl-free | pictures plus text | Good (14 AA v5) | opencode | 8.0 s |
| 17 | Nemotron 3 Ultra 550B nvidia/nemotron-3-ultra-550b-a55b | long reasoning and agents | Good (12 AA v5) | nvidia | 2.0 s |
| 18 | Gemini 3.5 Flash-Lite gemini-3.5-flash-lite | cheap fast short answers | Good (12 AA v5) | gemini | 3.2 s |
| 19 | Nemotron 3.5 Lightning 30B nvidia/nemotron-3.5-lightning-30b-a3b | fast cheap short answers | Small & fast (6 AA v5) | nvidia | 1.0 s |
| 20 | Nemotron 3.5 Lightning (free) nemotron-3.5-lightning-free | fast cheap short answers | Small & fast (6 AA v5) | opencode | 11.8 s |
| 21 | Muse Glimmer 30B meta/muse-glimmer-30b | fast light chat | Small & fast (6 AA v5) | nvidia | 6.2 s |
| 22 | ALLaM 2 7B allam-2-7b | Arabic and English chat | Small & fast | groq | 0.2 s |
| 23 | Gemini Flash-Lite (latest alias) gemini-flash-lite-latest | cheap fast short answers | Not rated yet | gemini | 0.6 s |
| 24 | Gemma 4 31B gemma-4-31b-it | text, pictures, general chat | Not rated yet | gemini | 28.0 s |
| 25 | Gemma 4 26B A4B gemma-4-26b-a4b-it | fast text and pictures chat | Not rated yet | gemini | 1.9 s |
| 26 | Nemotron 3 Nano Omni 30B nvidia/nemotron-3-nano-omni-30b-a3b-reasoning | text, pictures, audio understanding | Not rated yet | nvidia | 1.4 s |
| 27 | Nemotron 3.5 Bolt E4B nvidia/nemotron-3.5-bolt-e4b | tiny fast simple tasks | Not rated yet | nvidia | 0.2 s |
| 28 | Big Pickle big-pickle | coding (stealth test model) | Not rated yet | opencode | 9.5 s |
| 29 | Flint 40B flint-40b-free | unknown, test model | Not rated yet | opencode | 7.7 s |
| 30 | Gemini Robotics-ER 2 Preview gemini-robotics-er-2-preview | robot planning and pictures | Not rated yet | gemini | 1.4 s |
I read our notes, work logs and handovers to find the problems that keep coming back. Each one has a pain score, what we tried, and the free tools in this issue that could help.
A castle goes from picture to TRELLIS or Pixal3D to Blender and scores about 55 to 61 out of 100. Each new version helps a little, then the next one is inside the noise.
How often: 27+ castle versions since 2 Nov; best 61.2 paint-only (v27); several versions wasted (v14/v15 camera setting never applied, v18 three changes at once, v24 out of memory).
What we tried: One-change-per-version loop, 58-check form, noise band, Pixal3D vs TRELLIS, new camera fit, auto rating script.
Units shiver near walls, gates and forts. Catapults jam at their own wall. Each fix moves the problem somewhere else.
How often: At least 4 rounds in 4 days (#369, #382, #392, #391); gate crowd jitter work still open on 3 Nov; units freezing work open.
What we tried: jitter-detect, jitter-lab, jitter-soak tools; rule tests GR-UNIT-60/61; path grid and breach costs.
Tripo makes good-looking soldiers but the skeleton and mesh are not game-ready: no finger bones, twisted joints, seams. Only a few animation clips work on the newest rig.
How often: Reported in the Oct 28 research; Tripo rig v3.0 has 11 clips only; every new unit needs the same care.
What we tried: Master Body in Blender with one Rigify skeleton, Quaternius CC0 animations, Tripo pre-flight checklist (60 cr model, 25 cr rig).
Only 12 sound files play. 57 named sounds (unit appears, dies, catapult, gate, castle hit) have no file. We picked the free sound maker MOSS but making the real sounds is parked.
How often: Audit on 0.16.333: 11 of 12 files play, 5 made in code, 57 silent, 7 voice lines missing.
What we tried: MOSS-SoundEffect v2.0 installed (14.9 GB VRAM, 16-24 s per sound), 15 samples made; Kenney free sounds; own synth.
The audit found AI wasted half its words, ignored BUY and EARN, and picked units by turn order. Hard barely beats Medium. Single-player games mostly end at the 20 min limit.
How often: Audit 1 Nov; AI gates rewritten 3 times (GR-AI-29 then 32); parity law covers 8+ PRs; Hard vs Medium work open.
What we tried: Parity probe, soak runs, gate and wall rules, group formations.
A full test run takes about 15 min on the 3090. Some browser tests fail alone and pass later. Sim tests hang on the stacked build, and one test fails only with another change.
How often: Several named flaky tests in the handover; #440 hang and #445 clash open on 3-4 Nov.
What we tried: Own-PC rig, SwiftShader fallback, GPU go/stop lights, stack branches, no GitHub Actions.
Free model names change, quotas run out by mid-morning, and Cloudflare blocks some calls. On the first ping only 29 of 120 models answered.
How often: Daily; judge models moved on 3 Nov; Gemini free quota gone by late morning.
What we tried: Daily model ping task, WORKING-MODELS.md, fallback chain Gemini then Groq.
The game has no background music. ACE-Step made sad or noisy tracks; 384 s single songs sounded terrible. Only 190 s clips are clean.
How often: Two rounds in Oct 30 (5 loops rejected for minor and sad, 10 long songs rejected).
What we tried: ACE-Step 1.5 XL, shorter 190 s clips, crossfade idea, strict mood brief (calm, major, not sad).
Online play needs lockstep sync, rejoin after a drop, pause limits and swipe guards. Each new rule adds more cases. Livetest data is still in the live database.
How often: Lobby, offline mode, 4-player rooms, rejoin and pause plans in about a week.
What we tried: Fake server and bot phone for tests, AI takeover design.
Three looks, then three more looks, were drawn; Sandis rejected the second set. Then we switched to live 3D only. Units, castles and buttons must all match.
How often: Two look books in 1 day, then a 'newest way only' decision on 2 Nov.
What we tried: Z-Image Turbo, BiRefNet, Blender toon paint, style PDFs, score cards.
Files copied between PCs can be stale, rating scripts waited on the wrong file, the lead cannot message the Mini Boss, and the lead refuses work messages. Metrics were forgotten version after version.
How often: Stale copy seen 3 Nov; unrated v26/v27 for an hour; 24-attachment limit hit.
What we tried: New file names per commit, coded check form, job board, queue law.
The Play listing pack exists as a draft but store assets are missing, and Google Play and Apple work is parked. Data-delete route and privacy rules had gaps.
How often: Draft written 28 Oct; Google Play, Apple and old phones all on the parked list.
What we tried: Listing pack, Android plan, store-readiness checklist.
Sir Rowan's tutorial uses Gemini voice 'Orus'. 7 of 113 lines have no file. The new Sir Rowan voice work was parked.
How often: Audit of 30 Oct plus park note 3 Nov.
What we tried: Gemini TTS, one mp3 per line.
Thinking and coding AI that we use online with free keys. Each free plan has daily limits, and some may read our prompts, so the catch is printed on every card. Plans that only give a short trial live in the last chapter.
A 27B model that writes code, solves problems and reads pictures. We use it online through our free Groq key.
What's new: Released Aug 2026; vendor says SWE-bench Pro 61.7, Terminal-Bench 2.1 73.0, LiveCodeBench v6 90.3 (up from Qwen3.6-27B).
People say: Vendor numbers only; Groq and Cerebras also host it free (see api items).
My verdict: Top free online coding brain with no license strings, and very fast on Groq.
Free keys for Gemini 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite, plus Live and Transcribe models.
What's new: Gemini 3.8 Flash is free of charge on the pricing page.
My verdict: Best free hosted brain, if we mind the data rule.
Very fast free API for gpt-oss-120b, gpt-oss-20b, Qwen3.8-27B, Whisper.
What's new: Free limits: 30 requests/min, 1,000/day, 8K tokens/min, 200K tokens/day for the LLMs; Whisper 20 req/min, 2,000/day.
My verdict: Fast and free for dev work.
Free: Big Pickle, Pulse Alpha, GLM-5.3-Flash, DeepSeek V4.1 Flash, Beam, Flint 40B, MiMo-V2.6-Flash, Nemotron 3 Ultra and 3.5, Jev 1.13.
What's new: Docs list 11 free models.
My verdict: Use the zero-retention ones for our code.
Huge open coding/reasoning model (552B backbone, only 8-16B used per word). Too big to download for us, but free to call online.
What's new: Sept 10 2026 per aggregator. DeepSeek reports Terminal-Bench 2.1 90.6, DeepSWE 74.2, GPQA 90.9. Free on OpenCode Zen and build.nvidia.com.
People say: Artificial Analysis Intelligence Index about 33-45 depending on source/setting.
My verdict: Strongest free coder we can reach through OpenCode.
Google's open Gemma 4 models. We use the 31B and the fast 26B-A4B online through our free Gemini key.
What's new: Card shows technical report published Jul 2 (2026). Card numbers: LiveCodeBench v6 80.0, GPQA Diamond 84.3, AIME 2026 89.2.
People say: Also free on Gemini API and OpenRouter (gemma-4-31b-it:free).
My verdict: Apache-2.0 Gemma is a safe, fast second brain.
OpenAI's open-weight reasoning models. the 120B runs free online.
What's new: Offered free on Groq, Cerebras trial, and Cloudflare Workers AI right now.
My verdict: Reliable, widely hosted free models.
10,000 Neurons a day free, resets 00:00 UTC. Includes gpt-oss-20b/120b, Llama, Whisper, Flux. Small free amount; good for in-game helpers later.
A 501B open coding model with a plain Apache license, built in the USA. Free via OpenCode, but we could not open its own model card.
320B mixture model (18B active), cheap and fast. Nice free backup coder with good privacy terms.
A 30B model that uses only 3B at a time, so it is fast. Free online on build.nvidia.com and OpenCode. Fast and free with an open, permissive license, but weaker than Qwen3.8-27B.
Models ending in :free, e.g. gemma-4-31b-it:free, nemotron-3-super-120b-a12b:free, laguna-s-2.1:free; rotates. Only 50 requests a day, and stealth models are risky.
Free tier with 'restrictive rate limits'. Numbers unknown.
4 more free brains tools are free but have license strings, so they live in the last chapter: Kimi K3, MiniMax M3, NVIDIA build.nvidia.com free endpoints, Cerebras free trial.
Our biggest pain. Free online 3D AI is still rare: most 3D makers want to be downloaded, and we do not do that. So this chapter has the online services, plus free ready-made 3D kits and the free rigging website.
Free CC0 PBR textures, HDRIs and 3D models.
My verdict: Best free texture source; download by hand, do not scrape.
Free low-poly 3D packs: medieval buildings, ruins, knights, animated characters, a retargetable animation library.
What's new: Medieval Village MegaKit and Universal Animation Library are listed on the site.
My verdict: CC0 and already low-poly: perfect baseline.
Free website that rigs a humanoid and gives hundreds of ready animations.
My verdict: Free and legal for games; do not rebuild a clip library for resale.
Free CC0 game asset kits, including castle and medieval 3D kits.
My verdict: CC0 placeholder art while AI castles improve.
Lets an AI assistant control Blender: make objects, materials, run Python, check views and renders.
What's new: Renamed: package is now mcp-for-blender; old setups still work. 33.2k stars.
My verdict: Free helper that can script repetitive Blender fixes.
3 more castles & characters tools are free but have license strings, so they live in the last chapter: Hugging Face ZeroGPU Spaces (called by script), Tencent Hunyuan 3D (global app + Tencent Cloud API), NVIDIA build.nvidia.com hosted models (TRELLIS and image models).
Pictures for concept art, icons, store graphics and the pictures we feed into 3D makers. Here are the free online picture services and free art packs.
About 4,000 free black-and-white game icons (swords, castles, shields).
My verdict: Perfect medieval icon set; just add credits.
You send a text prompt in an HTTP call. You get a picture back as base64. Free allowance resets every day.
What's new: Pricing page (checked 2026-11-04) now lists FLUX.2 dev, FLUX.2 klein 4B and 9B, and Leonardo models. Some models need a paid method, not free.
My verdict: Free, simple REST call, fits our Cloudflare setup.
Free game art packs: buttons, UI, icons.
My verdict: Zero-risk free filler art.
6 more pictures & icons tools are free but have license strings, so they live in the last chapter: Gemini API image models (Nano Banana 2, Gemini 3.1 Flash Image), Hugging Face Inference Providers, Pollinations.ai (gen.pollinations.ai), Stability AI API (SD 3.5, remove background, upscale, Stable Fast 3D), remove.bg API, fal.ai, Replicate, Lumenfall (credit-based clouds).
Voices, sound effects and music. Free online voices are the happy story. For sound effects, real free recordings still beat AI.
A free zip of pro sound effects, 7.47 GB+ in the 2026 bundle.
What's new: 2026 bundle announced 2026-03-16; older years also available.
My verdict: Best free pro SFX pack, read the licence PDF once.
A huge site of user sounds. Filter to CC0 and you can use them freely.
My verdict: Real recordings, free, safe if you filter CC0.
Turns text into speech over an HTTP call. Has 30 ready voices, including Orus. Works in 130+ languages.
What's new: Pricing page shows TTS free of charge on the free tier. Paid price goes up on 2027-01-01.
My verdict: We already have the key and the voice, and it costs nothing.
A free audio editor to cut, fade, loop and clean sounds.
What's new: Audacity 4 is in development.
My verdict: Free and easy for trimming and fixing sounds.
Turns audio into text very fast. Can also translate speech to English.
My verdict: We have the key. Fast and free for testing voice features.
Free UI click, impact and RPG sound packs.
My verdict: Easy UI sounds with no credit needed.
Text-to-speech in several languages, run on Cloudflare's servers. Returns MP3.
My verdict: Same place as our website, cheap in free allowance.
Speech-to-text, with timing (VTT subtitles). Runs on Cloudflare. Good backup if Groq limits hurt.
Expressive text-to-speech with 'vocal direction' tags, such as a whisper or a laugh. Only 100 calls a day, and terms are not clear.
6 more music, sounds & voices tools are free but have license strings, so they live in the last chapter: OpenGameArt (music and SFX), ElevenLabs Free plan (TTS, sound effects, music), Deepgram $200 free credit (Nova-3 speech-to-text, Aura-2 text-to-speech), AssemblyAI free tier (speech-to-text), Google Cloud Text-to-Speech free monthly characters, Gemini API Lyria 3.5 and Lyria 3 Clip (music).
Old, proven, free tools that hit our everyday problems: crowds jamming at gates, slow tests, big 3D files and shipping to the stores. These are plain programs, not AI models, plus helpers that let free online AI do jobs for us.
A map and path tool for the browser. It makes walkable maps and has a crowd mover that stops units bumping.
My verdict: Best fix for crowd jitter at gates; real crowd steering.
Shrinks 3D files. Can use Draco, Meshopt and KTX2 texture squeeze.
My verdict: Biggest phone-speed win for generated castles.
Wraps a web app into real Android and iOS apps.
My verdict: The standard way to ship a web game to stores.
Controls real browsers for tests, with auto-waiting, traces and phone emulation.
My verdict: Best way to test the real game, with timeouts so tests cannot hang.
Free tileable PBR ground and wall textures.
My verdict: Safe free textures.
Free terminal coding agent that works with many models, including free ones.
My verdict: Already in our kit.
Helper library for WebSocket rooms on Durable Objects; PartySocket reconnects clients.
My verdict: Direct help for our rejoin problem.
Bad-word lists in 28 languages.
My verdict: Fast language filter.
Makes picking and hit-testing on Three.js meshes much faster.
My verdict: Cheap speed win for tapping and ground checks.
Makes meshes smaller and faster to draw. Can reduce triangle count.
My verdict: Proven classic for lighter 3D.
Fast TypeScript test runner with timeouts and watch mode.
My verdict: Quick tests with timeouts.
Makes random test inputs and shrinks a failure to a small example.
My verdict: Finds the weird bugs we cannot think of.
A small JS library for game brains: steering, state machines, goals, fuzzy logic and navmesh.
My verdict: Easy brain parts for bots; works with Three.js.
Lets an AI agent drive a browser and look at the page.
My verdict: Gives agents eyes on the game.
Turns words into number lists so we can find similar words and meanings.
My verdict: Cheap in the free pool and good for word play ideas.
Checks text (and images) for hate, harassment, violence, sexual content and more. Free to call.
My verdict: Free and made for player-name and chat checks.
Translates text between many languages with one HTTP call.
My verdict: Free draft translations for UI text.
Terminal pair programmer that edits files and commits to git. Fine, but OpenCode already covers it.
Open agent with desktop app, CLI and 70+ MCP extensions. Good backup agent.
Coding agent for VS Code, JetBrains and CLI. Similar to OpenCode.
Node tool to make textures from scratch. Handy extra.
The original wave function collapse maps from tiles. Good for tile maps; check a JS port too.
Low-code event flows. Free automation option.
Makes fantasy maps in the browser. Better for worlds than small battle maps.
Free API plan to try Mistral models. Mistral also lists Voxtral speech-to-text, embeddings and moderation models. Our prompts may train their models, so we only use it for harmless text.
4 more the game workshop tools are free but have license strings, so they live in the last chapter: wooorm/dictionaries (Hunspell), wordfreq, Kaikki.org Wiktionary data, n8n.
Free GitHub projects and ready-made workflows. Some help with 3D files, some are agent kits that let free online models work step by step.
Checks if a glb/gltf file is correct and writes a report.
My verdict: Cheap safety net for asset QA.
Small agent library where the AI writes short Python to do tasks.
My verdict: Smallest and simplest, works with free/local models.
Big list of AI tools for games: image, 3D, audio, LLM, agents.
What's new: 1,199 commits.
My verdict: Bookmark it for scouting.
Lets you draw an AI job as steps and loops with a saved state. Powerful but heavy for small team.
Makes teams of AI roles (planner, coder, tester) that pass work along. Popular, but loops can burn tokens.
Write AI steps as code and let it tune the prompts itself. Great, but learning curve.
Self-hosted app to run coding agents in safe sandboxes. Big overlap with agents we already use.
A tiny coding agent in about 100 lines. Easy but only one more agent.
Short curated list of AI tools for game makers. Probably quiet; use only as backup.
Finding a tool is not enough. Each one gets a fair, safe test, and then a clear verdict: keep it or throw it out. AI is always called online, never downloaded. Nothing is tested on GitHub, and nothing costs money.
A tester asks each free model 5 real Word Push questions (word puzzles, unit stats, a bug hunt) and marks the answers by itself.
Ask the online service for one castle tower from our own picture, open it in Blender, check shape, piece count and paint.
Ask the online service for one Archer icon in our style at 512 x 512 and compare it with our style sheet.
One narrator line ("The gate is open!") from an online voice, and one "sword hit" from the free sound packs.
Try the tool on a small copy of the game code, never on the real code folder.
| Verdict | When |
|---|---|
| KEEP | Total 18 or more out of 25, helps us 3+, safety 4+, licence OK. |
| THROW OUT | Total 11 or less, or safety 2 or less, or it does not help, or it wants money. |
| RETEST | Total 12 to 17, or the test broke because of us or a busy server. |
| BLOCKED | The rules are unclear, ban Europe, ban selling, or it only works if downloaded. |
| # | Tool | Lane | Machine | Status |
|---|---|---|---|---|
| 1 | Qwen3.8-27B | brains | cloud | waiting |
| 2 | Gemini API free tier (AI Studio) | brains | cloud | waiting |
| 3 | Poly Haven | 3d | Assistant PC | waiting |
| 4 | Sonniss GDC Game Audio Bundle 2026 | sound | Assistant PC | waiting |
| 5 | recast-navigation-js (WebAssembly Recast + Detour) | workshop | Assistant PC | waiting |
| 6 | glTF Transform (CLI + SDK) | workshop | Factory PC | waiting |
| 7 | Capacitor | workshop | Factory PC | waiting |
| 8 | Playwright (test runner + library) | workshop | Factory PC | waiting |
| 9 | ambientCG textures | workshop | Assistant PC | waiting |
| 10 | Groq free plan | brains | cloud | waiting |
| 11 | OpenCode Zen free models | brains | Factory PC | waiting |
| 12 | Quaternius packs | 3d | Assistant PC | waiting |
Where the news comes from. I start with 40 sources about free online AI and game tools. Every day each one gets a little score: did it bring true, useful news? Good sources rise, noisy ones sink, and new ones get added.
| Source | Kind | Trust | Why we trust it |
|---|---|---|---|
| Hugging Face Blog | official | 5/5 | Hugging Face hosts almost every open model, and its blog explains the big new ones. |
| Google DeepMind Blog | official | 5/5 | It is Google's own AI lab writing about its own models. |
| Google Blog: AI | official | 5/5 | Official Google news about its AI products. |
| Gemini API Changelog | official | 5/5 | It is the official change list for the Gemini API we already use with a free key. |
| Mistral AI News | official | 5/5 | Mistral is a European AI company and posts its own model news. |
| Microsoft Research Blog | research | 5/5 | Microsoft Research made TRELLIS, which we use for castles. |
| Cloudflare Developer Changelog | official | 5/5 | Official changes for Cloudflare, where Word Push runs. |
| Groq Changelog | official | 5/5 | Official change list for Groq, where we have a free key. |
| three.js releases | repo-releases | 5/5 | Official repo of the 3D library Word Push is built on (MIT). |
| OpenCode releases | repo-releases | 5/5 | Official repo of the OpenCode coding agent we already use. |
| SPDX license list data | repo-releases | 5/5 | SPDX is the official list of license names used everywhere. |
| Blender.org news | official | 5/5 | Official Blender Foundation news. |
| Blender Developers Blog | official | 5/5 | Blender's own developers write here. |
| Poly Haven assets API | official | 5/5 | Every Poly Haven model, texture and HDRI is CC0 (fully free). |
The full list of 40 sources is in the .
What users and reviewers say about tools from earlier issues this week.
These online services and tools are free to use, but their rules have a catch. Some rules are fine for us, some are walls. Each card says what the rule is, quotes the licence, and gives my advice: yes, maybe, or no.
My advice for a small game studio in Europe that sells its game. This is not legal advice: for anything big, we check the licence with a real lawyer first.
Visual workflow automation.
Yes, use it
Note: Strings and Node-RED does the job.
✓ Checked against its sourceWebsite that makes 3D models from text or pictures.
Maybe, carefully
✓ Checked against its sourceWebsite that makes full songs from words.
No, avoid it
✓ Checked against its sourceWebsite for very natural voices and sound effects.
No, avoid it
✓ Checked against its sourceBig game engine, free to download.
No, avoid it
✓ Checked against its sourceThese came up in the searches for the other pages. They are good tools, but each has a license rule, so they wait here.
Giant 2.8T open model, top open score on Artificial Analysis (Kimi K3.1 Pro 46). Catch: Custom license. If a product has over 100 million monthly users or over 20 million USD monthly revenue, it must show "Kimi K3" in its screen. Selling it as a hosted model service over 20 million USD yearly needs a deal. No region ban. Internal use is exempt.
428B open coding model. Catch: Not "non-commercial only". Commercial use is allowed if you email a one-time notice to api@minimax.io and show "Built with MiniMax M3". Over 20 million USD yearly revenue you need written permission. No region ban.
45 free endpoints incl. DeepSeek V4.1 Flash, GLM-5.3-Flash, Mistral Large 4, MiniMax M3.1 Flash, Nemotron 3.5. Catch: Prototyping only. Reports say production needs NVIDIA AI Enterprise; OpenCode says NVIDIA free endpoints are logged and 'trial use only'. Limits not published (users report about 40 requests/min).
gpt-oss-120b and qwen-3.8-27b free, but only a trial. Catch: No permanent free tier: $5 credits after adding payment method, expire in 30 days. 5 requests/min.
Many people host free image and image-to-3D demos on Hugging Face Spaces. Gradio's client can call them from code. Free GPU time is shared. Catch: Free GPU time: unsigned 2 minutes a day, free account 5 minutes a day (PRO 40). Each Space has its own model licence, many are non-commercial. Whether script calls count the same as browser use: not confirmed. Spaces can break or disappear.
Makes 3D models from text or pictures. The web app is free for some daily runs. A paid-cloud API also exists. Catch: Tencent says 20 free generations a day in the Hunyuan 3D Global web app, but that is a website, not an API. The API is for 'global enterprises': 200 free credits (one-off) via Tencent Cloud, and the Tencent Cloud FAQ says free quota is a one-time pack that expires after 1 year. Needs a Tencent Cloud account (card requirement not confirmed). Commercial use of free output: not confirmed. EU/UK availability: not confirmed.
NVIDIA hosts models you can call by API with a free trial key. A TRELLIS page exists there. Catch: NVIDIA calls the catalog a 'trial experience' for evaluation and prototyping. Production use needs an NVIDIA AI Enterprise licence. Number of free credits: not confirmed. Whether the TRELLIS page allows free API calls: not confirmed.
Google's image models for text-to-image and image editing through the Gemini API. Catch: No free tier for the image models on the pricing page. Free tier column says 'Not available'. Only usable with a paid account.
One HTTP API that sends your request to many image and 3D hosts (fal, Replicate, and others). Catch: Free accounts get $0 monthly credits now. You must buy credits or pay for PRO ($2 monthly credits included). So it is not free.
OpenAI-style API for images and text. You call a URL with your prompt and get a picture back. Catch: A third-party page says the free hourly refill ended on 2026-07-10 and free Pollen now only comes once from setup quests (about 3 Pollen, about $3). I could not confirm this on an official page. Official docs say you need a GitHub login and an API key. Old 'pk_' keys are limited to 1 Pollen per IP per hour and should not be made any more. No output ownership text found.
REST API for images, edits, upscaling, background removal and quick 3D. Catch: New accounts get 25 free credits (one-off, 1 credit = $0.01). That is about 8 SD 3.5 Medium images. Then you pay. Output licence: not confirmed.
Removes picture backgrounds by API. Catch: First 50 API calls per month are free. Free output size and commercial use: not confirmed. The standalone website ends on 2026-12-01 and background removal moves to Canva.
Pay-per-use APIs for many image, video and 3D models. Catch: fal: no general free tier found. Its docs only mention up to $1,000 credits for selected creators who apply. Replicate: pricing page does not say if there is a free trial. Lumenfall: says 'free credits included' with no amount, no card to sign up. Free video APIs: I found none I could confirm from an official page.
Community site with free game music and sounds. Catch: Each file has its own licence. CC-BY and OGA-BY need credit; SA and GPL add rules. Pick CC0 for easy life.
Great voices, sound effects and music. The Free plan gives 10k credits per month. Catch: Free plan is NON-COMMERCIAL only and needs credit to ElevenLabs ('If you are on the free plan, you can use the content non-commercially with attribution'). Commercial license starts at the paid Starter plan. Free API access: not confirmed.
Speech-to-text and text-to-speech APIs. New accounts get $200 of credit. Catch: One-time credit, not a monthly free tier. The pricing page says it does not expire, no card to start. After credit ends it is pay-as-you-go. Commercial output terms: not confirmed.
Speech-to-text API. Page says free tier has up to 185 hours pre-recorded and 333 hours streaming. Catch: Hours may be a one-off trial: expiry not stated on the page. FAQ says no card. Free plan only 5 new streams per minute. Commercial terms: not confirmed.
Classic cloud TTS. Free part: WaveNet and Standard up to 4 million characters per month, Chirp 3 HD and others up to 1 million. Catch: Page says: 'You must enable billing to use Text-to-Speech, and will be automatically charged if your usage exceeds the number of free characters.' So it needs a billing account and a card. Gemini-TTS models have no free characters here.
Makes music from text. Clip model: 30 seconds. Lyria 3.5: full songs. MP3 or WAV, with a hidden SynthID mark. Catch: A benchmark site quotes Google as saying there is no free tier for Lyria: about $0.08 per song, $0.04 per clip. I could not read this on Google's own pricing page. Treat as NOT free.
92 spell-check dictionaries in many languages. Catch: Each dictionary differs: MIT, BSD, MPL, GPL, LGPL, AGPL (Hebrew), CC-BY-SA. GPL/AGPL dictionaries can force sharing; pick permissive ones.
Word commonness scores in 40+ languages. Catch: Some data CC BY-SA 4.0; credit for Google Ngrams and SUBTLEX. Snapshot to about 2021, unlikely updated.
Wiktionary words as JSON for hundreds of languages. Catch: Share-alike: if we ship a word list made from it, we must credit Wiktionary and share that list under CC BY-SA too. Kaikki page itself states no licence, only asks for a citation.
A fact-checker re-opened the source pages for 70 items and fixed 88 things before printing, including licence mistakes such as MiniMax M3 being allowed for selling (with a notice).
5 items could not be confirmed at their own source (Beam 1.0, Hunyuan 3D 3.0, the Sonniss bundle and Kaikki). They are marked "Not fully confirmed yet". Treat them as probably right, not proven.
Scores out of 5 are my judgement from the sources and user reports, not lab tests. The test bed will replace them with real results.
None of this is legal advice. Before anything goes into the shipped game, we read its licence once more.